Base44 Superagents-AI agent that does it all
Mini toolsBitVLA
By ustcwhy
BitVLA is a native 1-bit VLA policy where parameters are ternary {-1,0,1}, built on BitNet b1.58 and trained as a vision-language-action model; it also introduces Quantize-then-Distill to compress the vision encoder to 1.58-bit weights while aligning to a full-precision teacher, with reported memory and latency reductions while matching an OpenVLA-OFT baseline on tasks.
New
Multimodal
Gen 3
Overview
BitVLA is a 1-bit vision-language-action model for robotic manipulation designed to run efficiently on memory-constrained edge platforms.
Tools using BitVLA
No tools found for this model yet.
Last updated: March 4, 2026
